Defending against Prompt Injection with Structured Queries (StruQ) and Preference Optimization (SecAlign)

BAIR researchers propose two fine-tuning defenses against prompt injection — StruQ (structured query separation) and SecAlign (preference optimization) — that require no extra compute or human labeling. StruQ reduces optimization-free attack success rates to ~0%; SecAlign also limits strong optimization-based attacks to under 15%, cutting previous SOTA rates by 4x across 5 tested LLMs. Directly relevant for engineers building LLM-integrated applications where user documents, web retrieval, or external data are in the prompt context.